Research On Complete Reduction And Knowledge Extraction Through Variable Precision Rough Set And Its Application

Attribute reduction and knowledge extraction are the available approaches to data mining and knowledge discovery, and the variable precision rough set provides an important mathematical tool. In fact, a good knowledge system can reveal the valuable information hidden in an information system by integrating knowledge from complete reduction in a comprehensive manner. The thesis aims at studying the complete reduction with knowledge extraction theory and methods, and further applying them to brain cognition analysis. The research is supported by multiple National Natural Science Funds of China. Firstly, this thesis puts forward theoretically complete reduction and knowledge extraction using variable precision rough set models. And then complete attribute reduction algorithm and knowledge structure algorithm are proposed, and the knowledge structure is also carried out. Finally, the brain functional connectivities are analyzed from the brain cognition data by the proposed methods. These studies not only provide an effective approach to analyze brain cognition, but also are very helpful to the intelligent technologies such as like-brain. The thesis has mainly the following contributions:The variable precision rough set models are extended to provide a complete reduction and knowledge extraction theory, which includes the definitions of complete reduct, multi-knowledge and the indexes of systems completeness and three new concepts, namly distribution tables, genealogical binary trees and double-layer knowledge structure. Complete reducts, knowledge and their relationship are analized for knowledge space which is characterized by a double-layer structure in Hasse.A novel complete attribute reduction algorithm is also presented, in which a strategy from the genealogical binary tree structure is used to get all reducts. We introduce four important strategies, namely, distribution table abstracting, attribute rank dynamic updating, hierarchical binary classifying and genealogical tree pruning are proposed. These strategies contribute to improve the algorithm convergence. The completeness of our algorithm is proved theoretically and its superiority to existing methods for obtaining complete reducts is demonstrated experimentally.A multi-knowledge extraction framework is further provided according to the knowledge extraction, which help us span the gap effectively from data to reducts and then to knowledge. A parallel information entropy discretization algorithm is also proposed for data preprocessing. After obtaining the multi-knowledge system through a complete reduction set, we present the method of obtaining the decision values by using multi-knowledge system, as well as a knowledge structure constructing algorithm. Eventually, the information of the double-layers knowledge structure can be acquired through our proposed algorithm.The research of brain cognition confronts a great challenge, which is to fully extract useful knowledge from functional Magnetic Resonance Imaging (fMRI) data. Multi-knowledge extraction system which is developed according to multi-knowledge extraction framework provides strongly support for brain cognition analysis. Using the multi-knowledge extraction system, all reducts of fMRI data can be obtained and multi-knowledge from all reducts can be extracted, and then knowledge structure algorithm is used to build the double-layer structure of fMRI data. The multi-knowledge extraction system is concerned with the application of the integrated technologies of brain cognition, and it would provide an effective tool for brain cognition analysis.